How to prepare documents for AI

Text copied straight out of a PDF loses its headings, breaks mid-word at column edges, scrambles tables into a run of numbers, and carries no page numbers. An AI summary built on that text has nothing to cite; a scanned page has no text at all to copy.

A table with merged cells or multi-line headers is especially prone to this: numbers end up in the wrong row entirely, which is worse than losing formatting because it can look correct until someone checks it against the original.

Document Extractor returns plain text immediately using pypdf, then a layout-aware Markdown version shortly after, built with Docling. Every page gets a page anchor, headings form a clickable outline, and tables are detected so they can be exported as CSV or XLSX instead of retyped.

This two-step return, plain text first and Markdown a moment later, means a workflow does not have to wait on the slower, more thorough pass before it can start reading or searching the document.

A scanned or photographed page has no text layer, so extraction alone returns nothing for it. OCR recognizes those pages at 3 units each on Pro and API; the free plan reads only pages that already carry a text layer.

Because OCR is priced and run separately, a document with only a few scanned pages does not pay the OCR rate for the pages that already had a text layer.

Open the extracted Markdown in the Markdown Editor to fix a heading, clean a table or add a diagram, with live preview and nothing uploaded. Export the cleaned version as Markdown, DOCX or HTML once it is ready.

The same extracted text can be summarized, questioned with citations, translated, or chunked into RAG-ready JSON with offsets for a vector store, all from the same document without extracting it twice.

The Document Extractor's REST API returns the same plain text, Markdown, tables and chunked JSON, so a pipeline can run the identical extraction on every new file without touching the browser.